关键词: 半监督学习, 目标检测, 对比学习, 重利用不可靠伪标签, 端到端训练

Abstract: The key to semi-supervised object detection methods is to assign pseudo labels to the targets of unlabeled data. To guarantee the quality of pseudo-labels， the semi-supervised object detection methods usually use a confidence threshold to filter low-quality pseudo-labels， which will cause most pseudo-labels to be removed due to their low confidence. Contrastive learning is used to reuse most of low-confidence unreliable pseudo labels for boosting the performance of semi-supervised object detection method. Specifically， the pseudo-labels are divided into reliable and unreliable ones according to the prediction confidence. Besides the reliable pseudo-labels， the unreliable pseudo-labels are exploited as negative samples for model training of contrast learning. To balance the number of unreliable pseudo-labels between different classes， a memory module is designed to store the unreliable pseudo-labels of different batches in the training process. The experimental results show that the mAP of the improved semi-supervised method on COCO data set is 13.6%， 23.0%， and 27.5% with the labeling ratio of 1%， 5%， and 10%， which is better than the existing semi-supervised learning methods. On the COCO-additional data set， the mAP of the improved semi-supervised method reaches 44.7%， which is 4.5 percentage points higher than supervised learning.

Key words: semi-supervised learning, object detection, contrastive learning, reusing unreliable pseudo-labels, end-to-end training